ABSTRACT
Purpose Basket trials test the activity of drugs in multiple cancer types (tissues of origin or genotypes). Rigorous comparison across types is complicated by small numbers of patients and the lack of a control arm, motivating development of new analytical approaches, particularly as patient stratification becomes more common and refined.
Experimental Design We reanalyze published basket trials of neratinib in ERBB-mutant cancers and larotrectinib in TRK-fusion cancers, using Monte Carlo permutation tests in which an implicit “no response to therapy” null hypothesis is replaced with an empirically derived null of ‘no difference in response by tumor type” (or class of mutation). All enrolled patients contribute to null distributions for the analysis of therapy-associated volume change and Progression Free Survival (PFS).
Results Testing neratinib responses in the SUMMIT trial against a no difference null provides insights not obtainable using a conventional dichotomous assessment of volume changes. For example, breast cancers pass the dichotomous standard and exceed the no difference test for volume changes but not for PFS. Conversely, lung cancers fail the dichotomous test but exceed no difference tests for volume changes and PFS (P= 0.04 and P=0.003) and lung cancers are the sole type for which a specific genotype, ERBB2 Exon 20 mutation (P=0.01), is significantly associated with increased PFS.
Conclusions Monte Carlo permutation tests enable rigorous determination of tumor types most likely to benefit from therapy in basket trials. Reanalysis of data from SUMMIT identifies an overlooked therapeutic opportunity for neratinib in lung cancers carrying ERBB2 Exon 20 mutations.
Statement of Translational Relevance Basket clinical trials are increasingly common in oncology as a means to identify responsive cohorts in patient populations comprising tumors that arise from many different tissues, particularly when searching for genotypes predictive of outcome. Basket trials typically lack formal control arms and enroll multiple tumor types, each represented by a small number of patients. To overcome the inherent statistical challenges in such patient groups we propose a new biostatistical approach that uses empirical P values to test for differences in response across a population. The approach brings added rigor to the interpretation of basket trials, can be applied to both volume changes and PFS duration, and is potentially applicable to any trial in which subdivision of patient populations is desirable as a means to identify outliers or discover and test genetic biomarkers. Our approach is instantiated in open-source code making it simple for others to validate the method and test it on their own data.